Tessera (`tessera`)¶

Quick Facts¶

Field	Value
Model ID	`tessera`
Family / Source	GeoTessera precomputed embeddings
Adapter type	`precomputed`
Training alignment	N/A (precomputed product)

Tessera In 30 Seconds

Tessera is a precomputed 10 m global embedding product distributed as local GeoTessera tiles — rs-embed does no model inference here; it resamples the tiles covering your ROI (nearest neighbour, each tile through its own UTM CRS) onto the common EPSG:3857 grid at 10 m and returns the (D,H,W) embedding grid on that grid — the same grid provider-backed models sample on.

In rs-embed, its most important characteristics are:

output lives on the common EPSG:3857 grid at 10 m, pixel-aligned with provider-backed models for the same ROI: see Output Semantics
a ROI straddling a UTM zone boundary is served (tiles from both zones are resampled independently) with a UserWarning about the seam: see Preprocessing / Retrieval Pipeline
year-selector temporal semantics: TemporalSpec.range(...) silently uses the start year rather than doing real temporal filtering: see Retrieval Contract

Retrieval Contract¶

Field	Value
Backend	`auto` (legacy `local` still accepted)
`SpatialSpec`	`BBox` / `PointBuffer`, mapped to the common EPSG:3857 grid exactly as provider requests are
`TemporalSpec`	`year(YYYY)` — uses `.year`; `range(start, end)` falls back to `start`'s year; default year `2021`
Source	GeoTessera precomputed tiles
Product CRS	tile-native UTM (varies by tile); output resampled onto EPSG:3857
Product resolution	10 m
Cache directory	`RS_EMBED_TESSERA_CACHE`, or per-call `sensor.collection="cache:/path/to/cache"`
Side inputs	none

Temporal semantics

TemporalSpec.range(...) is a year selector here, not scene-level temporal filtering.

Preprocessing / Retrieval Pipeline¶

flowchart LR
    INPUT["SpatialSpec\n+ TemporalSpec"] --> GRIDDEF["ROI → common grid\n(EPSG:3857 @ 10 m)"]
    GRIDDEF --> QUERY["grid footprint\n→ query tile blocks"]
    QUERY --> RESAMPLE["Fetch tiles → nearest-neighbour\nresample onto the grid\n(UTM seam → warning)"]
    RESAMPLE --> POOL["pooled: vector"]
    RESAMPLE --> GRID["grid: (D,H,W)\non the common grid"]

Architecture Concept¶

flowchart LR
    subgraph "Precomputed Tiled Product"
        RES["10 m per pixel"]
        STORE["GeoTessera tile blocks\n(local cache)"]
    end
    subgraph Retrieval
        RES --> BBOX["ROI → common grid\nfootprint"]
        STORE --> TILES["Query + fetch\ntile blocks"]
        BBOX --> TILES
        TILES --> MOS["Per-tile nearest\nresample onto grid"]
    end
    subgraph Output
        MOS --> POOL["pooled: spatial\npooling over grid"]
        MOS --> GRID["grid: (D,H,W)\non EPSG:3857"]
    end

Environment Variables / Tuning Knobs¶

Env var	Default	Effect
`RS_EMBED_TESSERA_CACHE`	unset (GeoTessera default)	Local GeoTessera cache directory
`RS_EMBED_TESSERA_BATCH_WORKERS`	`4`	Batch worker count for `get_embeddings_batch(...)`

Non-env override

sensor.collection="cache:/path/to/cache" overrides the cache directory for one call.

Output Semantics¶

pooled: spatial pooling over the cropped embedding grid.

grid: (D,H,W) on the common grid. Metadata records input_crs=EPSG:4326, output_crs=EPSG:3857, scale_m=10, the north-up transform (affine, pixel → EPSG:3857), grid_hw, tile_crs (the UTM CRSs the tiles came from), resampling=nearest, and coverage (fraction of grid pixels a tile covered; the rest are zero vectors).

Examples¶

Minimal example¶

from rs_embed import get_embedding, PointBuffer, TemporalSpec, OutputSpec

emb = get_embedding(
    "tessera",
    spatial=PointBuffer(lon=121.5, lat=31.2, buffer_m=5000),
    temporal=TemporalSpec.year(2021),
    output=OutputSpec.pooled(),
    backend="auto",
)

Example cache override¶

# Example (shell):
export RS_EMBED_TESSERA_CACHE=/data/geotessera

Paper & Links¶

Publication: CVPR 2026

Reference¶

Tiles are placed through their own CRS/affine, so mixed UTM zones and rotated tiles are all handled; tiles must share the embedding dimension.
A ROI across a UTM zone boundary warns once per request: the two sides are resampled independently, so expect a seam along the boundary.
"No tiles found" usually means the ROI/year combination has no coverage in the GeoTessera cache, not that the cache is broken.

Tessera (tessera)¶